A 64-Bit Symbolic AI Design Puts Logic and Traceable Reasoning First
By George Harvey 651 views
A proposed 64-bit language system encodes categories and context for fast analogy making, offering a different path from pattern matching.



A computer that can show how it reached an answer poses a different challenge from one that produces convincing language. George Harvey says he spent years designing a system for that first goal: represent words as structured 64-bit values, then use ordinary logical operations to compare and combine them. His experiment offers a concrete alternative to language systems built around pattern matching, though it was never developed into a commercial product.
Harvey began the work around 1986, initially as a spare-time project. He had little interest in treating the Turing test as proof of intelligence: ELIZA, an early program that mimicked conversation, already showed how a machine could appear human without demonstrating a deeper grasp of meaning. He wanted a computer to reason in steps that a person could inspect, and to explain how it arrived at a conclusion.
His proposed route was what he called Artificial Cognition. Instead of treating words as strings to be processed indirectly, the system would map them into compact computer-native representations. In principle, those representations could be combined into sentences, translated into and out of human languages, and manipulated as structured ideas. A computer might, for example, process a work such as Shakespeare’s Henry V in its own representation before producing another language version. Harvey expected some nuance to disappear, as it can in human translation.
Words encoded as structured values
The design went through several iterations. Harvey first considered very wide processors, with registers of 256 bits or more, but later settled on 64-bit values that could run on standard computer hardware. He had considered whether a new processor instruction set might be necessary; he says the final approach did not require one.
Each word or phrase would occupy a 64-bit register. The representation could then be handled with familiar logical operations such as AND, OR and XOR. Harvey says he eventually defined a vocabulary of a little over 10,000 entries. A patent was issued for the design more than 20 years ago; he believes the invention is now in the public domain.
To build the vocabulary, Harvey spent two years defining entries and testing how they compared. When he had about 2,500, he wrote a program that accepted two English words or phrases, converted them into the system’s representations, and searched the vocabulary for other pairs with the same encoded relationship.
The search sounds expensive if every entry must be compared with every other one. Harvey says the structure allowed many candidates to be rejected quickly: most pairs were incompatible, and whole groups could be ruled out together. If the input pair was “train” and “airplane,” for instance, entries for colors or clothing could be excluded as categories.
In his account, the computer could identify a relationship through a small number of operations. An XOR between two encoded values produces their difference, or a fault if the values cannot be compared. Applying that difference to another dictionary entry can produce a corresponding value. If the second operation succeeds, the system has found an analogy; if it fails, the search continues.

Harvey tested the method with the pair “car” and “small passenger plane.” The results included “commuter train” and “commuter plane,” “cargo train” and “cargo plane,” and “ambulance” and “air ambulance.” Other matches included “main battle tank” and “heavy bomber,” “scout car” and “scout plane,” “race car” and “racing plane,” and “taxi” and “small commercial passenger plane.”
One result surprised him: “lion” matched with “eagle.” Other predators in the small test vocabulary did not form the same pair. Harvey says the encoded definitions included a contextual association linking both animals to authority, which made the match possible. The result illustrates the design’s central claim: a computer can work with relationships recorded in its representation, rather than infer every connection from repeated examples.
What the 64 bits describe
The 64-bit word is divided into fields that record different kinds of information. Harvey describes a hierarchy of categories: eight bits can represent 256 category values, arranged through supersets, groups and subsets. One broad category, “agent,” covers things that act. Its groups include living things and mechanisms, with more specific divisions such as animals and plants.
A separate category covers location and time. Within it, geography can be divided into subsets such as air, land, water and space. The first two fields Harvey describes, called BAS and ALT, can distinguish broad classes. A car, train, taxi, bus and bicycle can share the “land vehicle” classification; an airplane, blimp or kite can be assigned to “air vehicle.” In the same scheme, a lion, tiger, bear or dog can be a “land animal,” while a bat can be an “air animal.” The categories describe where something functions, not necessarily its biological classification.
These two fields use 16 bits, allowing 65,536 combinations. That is only part of the representation: additional fields carry other attributes. A source field, SRC, might indicate what powers a vehicle. A car could be marked as using chemical fuel or electricity; a sail-powered vehicle could be associated with air. A value of zero can mean that the entry leaves the property open, or that any applicable value will do.
Another field, MTD, describes how something operates. For a car, it might encode an engine or motor; a sail-powered vehicle would need a different value, while a buggy could be associated with an animal. A DST field can record a destination or purpose—for example, that a car carries people. A CXT field holds context, including the “authority” association Harvey says accounted for the lion-and-eagle result.
Harvey also proposed eight negation bits, each linked to one of the fields. The categories have fuzzy edges: a word’s meaning does not always fit a single clean box. But the system’s operations rely on strict rules about how defined values compare. In practice, the rules would need to reflect the part of the vocabulary being processed.
That distinction matters. The representation can store chosen associations, but those choices still need to be made and maintained. Harvey describes the framework as capable of holding far more potential word values than a human language needs. The challenge is less about reserving enough bit patterns than about building definitions that capture the distinctions a useful system needs.
Speed is only one part of the claim
Harvey says the work ran on a 32-bit computer that simulated the 64-bit machines then approaching the market. On a 64-bit processor, he estimated that a comparison would take one clock cycle and that the two operations needed for an analogy would take two. At 2 GHz, he calculated that the comparisons could be completed in one billionth of a second, including on a Raspberry Pi he priced at $45.
That figure describes the operations under the design’s assumptions; it does not establish the time required to build or search a complete, practical knowledge system. The vocabulary took years to assemble, and its usefulness depends on how entries are defined and on the rules applied to them. Fast comparison can make a particular operation cheap, but it cannot by itself supply reliable knowledge or settle ambiguous meanings.
Harvey says he hoped software could eventually build on the vocabulary by reading dictionaries. He also says he could not find anyone willing to evaluate the invention, and funding ended. The design therefore remained an experiment rather than a tested commercial system. Its proposed capabilities should be understood as Harvey’s account of what the architecture could do, not as results from a deployed product.
A different view of AI risk
Harvey sets his approach against current AI products, which he sees as relying heavily on pattern matching. His concern is that a system that imitates patterns in human thought can also reproduce human inconsistency. For him, an answer that sounds plausible does not demonstrate rational reasoning. A system should make its steps available for inspection.
He raises that concern in the context of warnings by Anthropic about potential harms from AI. The company has described severe risks in materials related to its planned stock offering, and Harvey questions the wisdom of treating systems with those risks as valuable products. His criticism is pointed: he argues that the field needs more attention to reasoning, caution and ethics, rather than assuming that stronger pattern matching will supply them.
The comparison has limits. Harvey’s system was designed for words and symbolic relationships; he explicitly says it does not handle images or other tasks that depend on pattern recognition. He left those applications to other researchers. His proposal, then, is not a general replacement for every current AI method. It is an argument for building systems whose representations and inferences can be examined.
Whether that approach can scale beyond a hand-built vocabulary remains an open question. Its lasting contribution may be the sharper standard it sets: when a computer gives an answer, can a person trace the categories, assumptions and operations behind it? Speed matters, but for systems used to make decisions, the ability to inspect their reasoning may matter just as much.
Featured image: “Fuzzy definition,” by Luke Jones, via Unsplash.